Policy at the Junction: The Unified Model Router as a Narrow Waist

A local program that uses several model providers must otherwise manage their different wire formats, credentials, conversation histories, failure modes and data policies. This working paper describes the Unified Model Router, a loopback service that speaks OpenAI's Responses API to local clients and places one protocol and one policy point between them and native OpenAI, Anthropic's Messages API and OpenRouter.The router is read as a narrow waist, drawn as a synapse: clients above, providers below, and a junction that decides what may cross. At that junction a deployer can declare a deny data policy, keep training-tier and protected sessions apart, pin OpenRouter endpoints, keep Claude on Anthropic's API, set a thinking budget where the model accepts one, require a source revision, and ask which route and provider served a call. Failures become named outcomes rather than a silent change of model. Throughout, the paper keeps apart what the router enforces, what it can only request of a provider, and what it can only report.It also sketches a proposed evidence gate, in which a host releases an action only when a Gadfly reading shows that the decision behind it rests on its declared evidence and the host's own policy permits it; the gate is a design, not an implemented feature. It describes the repository's review practice, in which Codex drives review models through the router and each finding becomes a guard test. The evaluation is an inventory of the offline test suite at commit 3f57e3d (176 tests, all passing), not a measurement of performance, provider compatibility or security. An appendix reads the paper's own citations against their abstracts with Gadfly, including a planted false claim as a control.The claim ledger (CLAIMS.md) ties each factual claim about the router to files and lines at commit 3f57e3d.